llama.cpp Releases

This release focuses on knowledge base management, wiki improvements, and MCP (Model Context Protocol) server fixes. It includes features for wiki revision history, document parsing, and knowledge configuration.

Aug 7, 2026

This release adds functionality to save and load chunks in the multi-modal (mtmd) system. It includes new tests to verify the functionality.

Aug 6, 2026

This release fixes a bug where the `/cors-proxy` endpoint was returning empty responses.

Aug 6, 2026

This release fixes a bug in the model loader where quantized tensors with reshaped strides were not handled correctly.

Aug 6, 2026

This release prepares the CI pipeline for AMD ROCm support and fixes specific issues with integrated GPUs (gfx1151). It includes workarounds for coherence issues and skips specific tests that fail on the HIP backend.

Aug 6, 2026

This release fixes a bug in Vulkan submission batching that could lead to DeviceLost errors. It also adds debug tools to help diagnose the causes of these driver errors.

Aug 6, 2026

This release introduces a new graph ordering function (`ggml_build_forward_order`) to prevent stale inputs in multi-modal audio processing graphs. It addresses a bug where compute flags were applied too early, causing incorrect execution in specific branches.

Aug 6, 2026

This release hardens the server's file search functionality to handle Windows junctions and path normalization correctly. It also fixes UI caching issues and improves error reporting for unreadable directories.

Aug 5, 2026

This build focuses on test infrastructure improvements, specifically re-enabling the MiniMax M3 architecture test. It includes standard pre-built binaries for multiple platforms and operating systems.

Aug 5, 2026

Fixes the max_tiles setting in the converter for the Unlimited-OCR feature.

Aug 5, 2026

Updates grammar constraints to treat max repetition values of 2000 or higher as unbounded.

Aug 5, 2026

Enhances multi-modal processing by supporting multi-row batching for the DeepSeek OCR model.

Aug 5, 2026

Fixes a memory allocation bug affecting MTP layers (Multi-Token Prediction).

Aug 5, 2026

Adds spec-decode counters to the /metrics endpoint to improve observability and align with vLLM standards.

Aug 5, 2026

Applies vendor-specific patches to fix issues with the subprocess.h header file.

Aug 5, 2026

Removes the GGML_METAL_USE_BF16 build flag from all build scripts.

Aug 5, 2026

Improves build security by switching the package installation command from npm install to npm ci.

Aug 5, 2026

Fixes a bug where accented characters were silently lost on Windows due to OEM code page encoding issues in built-in tools.

Aug 4, 2026

Corrects a bug involving duplicate empty audio chunks when processing short inputs in the multi-modal text-to-speech module.

Aug 4, 2026

Removes the 'full-context windows' parameter from history-based samplers.

Aug 4, 2026